1// Copyright 2016 the V8 project authors. All rights reserved. 2// Use of this source code is governed by a BSD-style license that can be 3// found in the LICENSE file. 4 5#include "src/compiler/memory-optimizer.h" 6 7#include "src/compiler/js-graph.h" 8#include "src/compiler/linkage.h" 9#include "src/compiler/node-matchers.h" 10#include "src/compiler/node-properties.h" 11#include "src/compiler/node.h" 12#include "src/compiler/simplified-operator.h" 13 14namespace v8 { 15namespace internal { 16namespace compiler { 17 18MemoryOptimizer::MemoryOptimizer(JSGraph* jsgraph, Zone* zone) 19 : jsgraph_(jsgraph), 20 empty_state_(AllocationState::Empty(zone)), 21 pending_(zone), 22 tokens_(zone), 23 zone_(zone) {} 24 25void MemoryOptimizer::Optimize() { 26 EnqueueUses(graph()->start(), empty_state()); 27 while (!tokens_.empty()) { 28 Token const token = tokens_.front(); 29 tokens_.pop(); 30 VisitNode(token.node, token.state); 31 } 32 DCHECK(pending_.empty()); 33 DCHECK(tokens_.empty()); 34} 35 36MemoryOptimizer::AllocationGroup::AllocationGroup(Node* node, 37 PretenureFlag pretenure, 38 Zone* zone) 39 : node_ids_(zone), pretenure_(pretenure), size_(nullptr) { 40 node_ids_.insert(node->id()); 41} 42 43MemoryOptimizer::AllocationGroup::AllocationGroup(Node* node, 44 PretenureFlag pretenure, 45 Node* size, Zone* zone) 46 : node_ids_(zone), pretenure_(pretenure), size_(size) { 47 node_ids_.insert(node->id()); 48} 49 50void MemoryOptimizer::AllocationGroup::Add(Node* node) { 51 node_ids_.insert(node->id()); 52} 53 54bool MemoryOptimizer::AllocationGroup::Contains(Node* node) const { 55 return node_ids_.find(node->id()) != node_ids_.end(); 56} 57 58MemoryOptimizer::AllocationState::AllocationState() 59 : group_(nullptr), size_(std::numeric_limits<int>::max()), top_(nullptr) {} 60 61MemoryOptimizer::AllocationState::AllocationState(AllocationGroup* group) 62 : group_(group), size_(std::numeric_limits<int>::max()), top_(nullptr) {} 63 64MemoryOptimizer::AllocationState::AllocationState(AllocationGroup* group, 65 int size, Node* top) 66 : group_(group), size_(size), top_(top) {} 67 68bool MemoryOptimizer::AllocationState::IsNewSpaceAllocation() const { 69 return group() && group()->IsNewSpaceAllocation(); 70} 71 72void MemoryOptimizer::VisitNode(Node* node, AllocationState const* state) { 73 DCHECK(!node->IsDead()); 74 DCHECK_LT(0, node->op()->EffectInputCount()); 75 switch (node->opcode()) { 76 case IrOpcode::kAllocate: 77 return VisitAllocate(node, state); 78 case IrOpcode::kCall: 79 return VisitCall(node, state); 80 case IrOpcode::kLoadElement: 81 return VisitLoadElement(node, state); 82 case IrOpcode::kLoadField: 83 return VisitLoadField(node, state); 84 case IrOpcode::kStoreElement: 85 return VisitStoreElement(node, state); 86 case IrOpcode::kStoreField: 87 return VisitStoreField(node, state); 88 case IrOpcode::kCheckedLoad: 89 case IrOpcode::kCheckedStore: 90 case IrOpcode::kDeoptimizeIf: 91 case IrOpcode::kDeoptimizeUnless: 92 case IrOpcode::kIfException: 93 case IrOpcode::kLoad: 94 case IrOpcode::kStore: 95 return VisitOtherEffect(node, state); 96 default: 97 break; 98 } 99 DCHECK_EQ(0, node->op()->EffectOutputCount()); 100} 101 102void MemoryOptimizer::VisitAllocate(Node* node, AllocationState const* state) { 103 DCHECK_EQ(IrOpcode::kAllocate, node->opcode()); 104 Node* value; 105 Node* size = node->InputAt(0); 106 Node* effect = node->InputAt(1); 107 Node* control = node->InputAt(2); 108 PretenureFlag pretenure = OpParameter<PretenureFlag>(node->op()); 109 110 // Determine the top/limit addresses. 111 Node* top_address = jsgraph()->ExternalConstant( 112 pretenure == NOT_TENURED 113 ? ExternalReference::new_space_allocation_top_address(isolate()) 114 : ExternalReference::old_space_allocation_top_address(isolate())); 115 Node* limit_address = jsgraph()->ExternalConstant( 116 pretenure == NOT_TENURED 117 ? ExternalReference::new_space_allocation_limit_address(isolate()) 118 : ExternalReference::old_space_allocation_limit_address(isolate())); 119 120 // Check if we can fold this allocation into a previous allocation represented 121 // by the incoming {state}. 122 Int32Matcher m(size); 123 if (m.HasValue() && m.Value() < Page::kMaxRegularHeapObjectSize) { 124 int32_t const object_size = m.Value(); 125 if (state->size() <= Page::kMaxRegularHeapObjectSize - object_size && 126 state->group()->pretenure() == pretenure) { 127 // We can fold this Allocate {node} into the allocation {group} 128 // represented by the given {state}. Compute the upper bound for 129 // the new {state}. 130 int32_t const state_size = state->size() + object_size; 131 132 // Update the reservation check to the actual maximum upper bound. 133 AllocationGroup* const group = state->group(); 134 if (OpParameter<int32_t>(group->size()) < state_size) { 135 NodeProperties::ChangeOp(group->size(), 136 common()->Int32Constant(state_size)); 137 } 138 139 // Update the allocation top with the new object allocation. 140 // TODO(bmeurer): Defer writing back top as much as possible. 141 Node* top = graph()->NewNode(machine()->IntAdd(), state->top(), 142 jsgraph()->IntPtrConstant(object_size)); 143 effect = graph()->NewNode( 144 machine()->Store(StoreRepresentation( 145 MachineType::PointerRepresentation(), kNoWriteBarrier)), 146 top_address, jsgraph()->IntPtrConstant(0), top, effect, control); 147 148 // Compute the effective inner allocated address. 149 value = graph()->NewNode( 150 machine()->BitcastWordToTagged(), 151 graph()->NewNode(machine()->IntAdd(), state->top(), 152 jsgraph()->IntPtrConstant(kHeapObjectTag))); 153 154 // Extend the allocation {group}. 155 group->Add(value); 156 state = AllocationState::Open(group, state_size, top, zone()); 157 } else { 158 // Setup a mutable reservation size node; will be patched as we fold 159 // additional allocations into this new group. 160 Node* size = graph()->NewNode(common()->Int32Constant(object_size)); 161 162 // Load allocation top and limit. 163 Node* top = effect = 164 graph()->NewNode(machine()->Load(MachineType::Pointer()), top_address, 165 jsgraph()->IntPtrConstant(0), effect, control); 166 Node* limit = effect = graph()->NewNode( 167 machine()->Load(MachineType::Pointer()), limit_address, 168 jsgraph()->IntPtrConstant(0), effect, control); 169 170 // Check if we need to collect garbage before we can start bump pointer 171 // allocation (always done for folded allocations). 172 Node* check = graph()->NewNode( 173 machine()->UintLessThan(), 174 graph()->NewNode( 175 machine()->IntAdd(), top, 176 machine()->Is64() 177 ? graph()->NewNode(machine()->ChangeInt32ToInt64(), size) 178 : size), 179 limit); 180 Node* branch = 181 graph()->NewNode(common()->Branch(BranchHint::kTrue), check, control); 182 183 Node* if_true = graph()->NewNode(common()->IfTrue(), branch); 184 Node* etrue = effect; 185 Node* vtrue = top; 186 187 Node* if_false = graph()->NewNode(common()->IfFalse(), branch); 188 Node* efalse = effect; 189 Node* vfalse; 190 { 191 Node* target = pretenure == NOT_TENURED 192 ? jsgraph()->AllocateInNewSpaceStubConstant() 193 : jsgraph()->AllocateInOldSpaceStubConstant(); 194 if (!allocate_operator_.is_set()) { 195 CallDescriptor* descriptor = 196 Linkage::GetAllocateCallDescriptor(graph()->zone()); 197 allocate_operator_.set(common()->Call(descriptor)); 198 } 199 vfalse = efalse = graph()->NewNode(allocate_operator_.get(), target, 200 size, efalse, if_false); 201 vfalse = graph()->NewNode(machine()->IntSub(), vfalse, 202 jsgraph()->IntPtrConstant(kHeapObjectTag)); 203 } 204 205 control = graph()->NewNode(common()->Merge(2), if_true, if_false); 206 effect = graph()->NewNode(common()->EffectPhi(2), etrue, efalse, control); 207 value = graph()->NewNode( 208 common()->Phi(MachineType::PointerRepresentation(), 2), vtrue, vfalse, 209 control); 210 211 // Compute the new top and write it back. 212 top = graph()->NewNode(machine()->IntAdd(), value, 213 jsgraph()->IntPtrConstant(object_size)); 214 effect = graph()->NewNode( 215 machine()->Store(StoreRepresentation( 216 MachineType::PointerRepresentation(), kNoWriteBarrier)), 217 top_address, jsgraph()->IntPtrConstant(0), top, effect, control); 218 219 // Compute the initial object address. 220 value = graph()->NewNode( 221 machine()->BitcastWordToTagged(), 222 graph()->NewNode(machine()->IntAdd(), value, 223 jsgraph()->IntPtrConstant(kHeapObjectTag))); 224 225 // Start a new allocation group. 226 AllocationGroup* group = 227 new (zone()) AllocationGroup(value, pretenure, size, zone()); 228 state = AllocationState::Open(group, object_size, top, zone()); 229 } 230 } else { 231 // Load allocation top and limit. 232 Node* top = effect = 233 graph()->NewNode(machine()->Load(MachineType::Pointer()), top_address, 234 jsgraph()->IntPtrConstant(0), effect, control); 235 Node* limit = effect = 236 graph()->NewNode(machine()->Load(MachineType::Pointer()), limit_address, 237 jsgraph()->IntPtrConstant(0), effect, control); 238 239 // Compute the new top. 240 Node* new_top = graph()->NewNode( 241 machine()->IntAdd(), top, 242 machine()->Is64() 243 ? graph()->NewNode(machine()->ChangeInt32ToInt64(), size) 244 : size); 245 246 // Check if we can do bump pointer allocation here. 247 Node* check = graph()->NewNode(machine()->UintLessThan(), new_top, limit); 248 Node* branch = 249 graph()->NewNode(common()->Branch(BranchHint::kTrue), check, control); 250 251 Node* if_true = graph()->NewNode(common()->IfTrue(), branch); 252 Node* etrue = effect; 253 Node* vtrue; 254 { 255 etrue = graph()->NewNode( 256 machine()->Store(StoreRepresentation( 257 MachineType::PointerRepresentation(), kNoWriteBarrier)), 258 top_address, jsgraph()->IntPtrConstant(0), new_top, etrue, if_true); 259 vtrue = graph()->NewNode( 260 machine()->BitcastWordToTagged(), 261 graph()->NewNode(machine()->IntAdd(), top, 262 jsgraph()->IntPtrConstant(kHeapObjectTag))); 263 } 264 265 Node* if_false = graph()->NewNode(common()->IfFalse(), branch); 266 Node* efalse = effect; 267 Node* vfalse; 268 { 269 Node* target = pretenure == NOT_TENURED 270 ? jsgraph()->AllocateInNewSpaceStubConstant() 271 : jsgraph()->AllocateInOldSpaceStubConstant(); 272 if (!allocate_operator_.is_set()) { 273 CallDescriptor* descriptor = 274 Linkage::GetAllocateCallDescriptor(graph()->zone()); 275 allocate_operator_.set(common()->Call(descriptor)); 276 } 277 vfalse = efalse = graph()->NewNode(allocate_operator_.get(), target, size, 278 efalse, if_false); 279 } 280 281 control = graph()->NewNode(common()->Merge(2), if_true, if_false); 282 effect = graph()->NewNode(common()->EffectPhi(2), etrue, efalse, control); 283 value = graph()->NewNode(common()->Phi(MachineRepresentation::kTagged, 2), 284 vtrue, vfalse, control); 285 286 // Create an unfoldable allocation group. 287 AllocationGroup* group = 288 new (zone()) AllocationGroup(value, pretenure, zone()); 289 state = AllocationState::Closed(group, zone()); 290 } 291 292 // Replace all effect uses of {node} with the {effect}, enqueue the 293 // effect uses for further processing, and replace all value uses of 294 // {node} with the {value}. 295 for (Edge edge : node->use_edges()) { 296 if (NodeProperties::IsEffectEdge(edge)) { 297 EnqueueUse(edge.from(), edge.index(), state); 298 edge.UpdateTo(effect); 299 } else { 300 DCHECK(NodeProperties::IsValueEdge(edge)); 301 edge.UpdateTo(value); 302 } 303 } 304 305 // Kill the {node} to make sure we don't leave dangling dead uses. 306 node->Kill(); 307} 308 309void MemoryOptimizer::VisitCall(Node* node, AllocationState const* state) { 310 DCHECK_EQ(IrOpcode::kCall, node->opcode()); 311 // If the call can allocate, we start with a fresh state. 312 if (!(CallDescriptorOf(node->op())->flags() & CallDescriptor::kNoAllocate)) { 313 state = empty_state(); 314 } 315 EnqueueUses(node, state); 316} 317 318void MemoryOptimizer::VisitLoadElement(Node* node, 319 AllocationState const* state) { 320 DCHECK_EQ(IrOpcode::kLoadElement, node->opcode()); 321 ElementAccess const& access = ElementAccessOf(node->op()); 322 Node* index = node->InputAt(1); 323 node->ReplaceInput(1, ComputeIndex(access, index)); 324 NodeProperties::ChangeOp(node, machine()->Load(access.machine_type)); 325 EnqueueUses(node, state); 326} 327 328void MemoryOptimizer::VisitLoadField(Node* node, AllocationState const* state) { 329 DCHECK_EQ(IrOpcode::kLoadField, node->opcode()); 330 FieldAccess const& access = FieldAccessOf(node->op()); 331 Node* offset = jsgraph()->IntPtrConstant(access.offset - access.tag()); 332 node->InsertInput(graph()->zone(), 1, offset); 333 NodeProperties::ChangeOp(node, machine()->Load(access.machine_type)); 334 EnqueueUses(node, state); 335} 336 337void MemoryOptimizer::VisitStoreElement(Node* node, 338 AllocationState const* state) { 339 DCHECK_EQ(IrOpcode::kStoreElement, node->opcode()); 340 ElementAccess const& access = ElementAccessOf(node->op()); 341 Node* object = node->InputAt(0); 342 Node* index = node->InputAt(1); 343 WriteBarrierKind write_barrier_kind = 344 ComputeWriteBarrierKind(object, state, access.write_barrier_kind); 345 node->ReplaceInput(1, ComputeIndex(access, index)); 346 NodeProperties::ChangeOp( 347 node, machine()->Store(StoreRepresentation( 348 access.machine_type.representation(), write_barrier_kind))); 349 EnqueueUses(node, state); 350} 351 352void MemoryOptimizer::VisitStoreField(Node* node, 353 AllocationState const* state) { 354 DCHECK_EQ(IrOpcode::kStoreField, node->opcode()); 355 FieldAccess const& access = FieldAccessOf(node->op()); 356 Node* object = node->InputAt(0); 357 WriteBarrierKind write_barrier_kind = 358 ComputeWriteBarrierKind(object, state, access.write_barrier_kind); 359 Node* offset = jsgraph()->IntPtrConstant(access.offset - access.tag()); 360 node->InsertInput(graph()->zone(), 1, offset); 361 NodeProperties::ChangeOp( 362 node, machine()->Store(StoreRepresentation( 363 access.machine_type.representation(), write_barrier_kind))); 364 EnqueueUses(node, state); 365} 366 367void MemoryOptimizer::VisitOtherEffect(Node* node, 368 AllocationState const* state) { 369 EnqueueUses(node, state); 370} 371 372Node* MemoryOptimizer::ComputeIndex(ElementAccess const& access, Node* key) { 373 Node* index = key; 374 int element_size_shift = 375 ElementSizeLog2Of(access.machine_type.representation()); 376 if (element_size_shift) { 377 index = graph()->NewNode(machine()->Word32Shl(), index, 378 jsgraph()->Int32Constant(element_size_shift)); 379 } 380 const int fixed_offset = access.header_size - access.tag(); 381 if (fixed_offset) { 382 index = graph()->NewNode(machine()->Int32Add(), index, 383 jsgraph()->Int32Constant(fixed_offset)); 384 } 385 if (machine()->Is64()) { 386 // TODO(turbofan): This is probably only correct for typed arrays, and only 387 // if the typed arrays are at most 2GiB in size, which happens to match 388 // exactly our current situation. 389 index = graph()->NewNode(machine()->ChangeUint32ToUint64(), index); 390 } 391 return index; 392} 393 394WriteBarrierKind MemoryOptimizer::ComputeWriteBarrierKind( 395 Node* object, AllocationState const* state, 396 WriteBarrierKind write_barrier_kind) { 397 if (state->IsNewSpaceAllocation() && state->group()->Contains(object)) { 398 write_barrier_kind = kNoWriteBarrier; 399 } 400 return write_barrier_kind; 401} 402 403MemoryOptimizer::AllocationState const* MemoryOptimizer::MergeStates( 404 AllocationStates const& states) { 405 // Check if all states are the same; or at least if all allocation 406 // states belong to the same allocation group. 407 AllocationState const* state = states.front(); 408 AllocationGroup* group = state->group(); 409 for (size_t i = 1; i < states.size(); ++i) { 410 if (states[i] != state) state = nullptr; 411 if (states[i]->group() != group) group = nullptr; 412 } 413 if (state == nullptr) { 414 if (group != nullptr) { 415 // We cannot fold any more allocations into this group, but we can still 416 // eliminate write barriers on stores to this group. 417 // TODO(bmeurer): We could potentially just create a Phi here to merge 418 // the various tops; but we need to pay special attention not to create 419 // an unschedulable graph. 420 state = AllocationState::Closed(group, zone()); 421 } else { 422 // The states are from different allocation groups. 423 state = empty_state(); 424 } 425 } 426 return state; 427} 428 429void MemoryOptimizer::EnqueueMerge(Node* node, int index, 430 AllocationState const* state) { 431 DCHECK_EQ(IrOpcode::kEffectPhi, node->opcode()); 432 int const input_count = node->InputCount() - 1; 433 DCHECK_LT(0, input_count); 434 Node* const control = node->InputAt(input_count); 435 if (control->opcode() == IrOpcode::kLoop) { 436 // For loops we always start with an empty state at the beginning. 437 if (index == 0) EnqueueUses(node, empty_state()); 438 } else { 439 DCHECK_EQ(IrOpcode::kMerge, control->opcode()); 440 // Check if we already know about this pending merge. 441 NodeId const id = node->id(); 442 auto it = pending_.find(id); 443 if (it == pending_.end()) { 444 // Insert a new pending merge. 445 it = pending_.insert(std::make_pair(id, AllocationStates(zone()))).first; 446 } 447 // Add the next input state. 448 it->second.push_back(state); 449 // Check if states for all inputs are available by now. 450 if (it->second.size() == static_cast<size_t>(input_count)) { 451 // All inputs to this effect merge are done, merge the states given all 452 // input constraints, drop the pending merge and enqueue uses of the 453 // EffectPhi {node}. 454 state = MergeStates(it->second); 455 EnqueueUses(node, state); 456 pending_.erase(it); 457 } 458 } 459} 460 461void MemoryOptimizer::EnqueueUses(Node* node, AllocationState const* state) { 462 for (Edge const edge : node->use_edges()) { 463 if (NodeProperties::IsEffectEdge(edge)) { 464 EnqueueUse(edge.from(), edge.index(), state); 465 } 466 } 467} 468 469void MemoryOptimizer::EnqueueUse(Node* node, int index, 470 AllocationState const* state) { 471 if (node->opcode() == IrOpcode::kEffectPhi) { 472 // An EffectPhi represents a merge of different effect chains, which 473 // needs special handling depending on whether the merge is part of a 474 // loop or just a normal control join. 475 EnqueueMerge(node, index, state); 476 } else { 477 Token token = {node, state}; 478 tokens_.push(token); 479 } 480} 481 482Graph* MemoryOptimizer::graph() const { return jsgraph()->graph(); } 483 484Isolate* MemoryOptimizer::isolate() const { return jsgraph()->isolate(); } 485 486CommonOperatorBuilder* MemoryOptimizer::common() const { 487 return jsgraph()->common(); 488} 489 490MachineOperatorBuilder* MemoryOptimizer::machine() const { 491 return jsgraph()->machine(); 492} 493 494} // namespace compiler 495} // namespace internal 496} // namespace v8 497